Anthropic’s latest security research reveals that traditional indicator-based cyberattack detection is failing as AI-driven threats prioritize complex behavioral sequences over simple, easily discarded technical artifacts.
Key Points
- Anthropic analyzed 832 malicious accounts and 13,873 observed actions between March 2025 and March 2026 to evaluate modern threat patterns.
- The study found that high-risk state-sponsored campaigns often use a similar number of techniques as low-risk actors, rendering technique-counting an ineffective security metric.
- AI-assisted cyberattacks are increasingly shifting toward post-compromise activities, with account discovery usage rising by 8.9% while phishing efforts declined.
- The proportion of medium-to-high-risk actors identified by Anthropic increased from 33% to 56% over the 12-month study period.
- Anthropic introduced the Adversarial Pattern Extraction and Correlation (APEX) framework to prioritize the sequence, timing, and context of attacks over individual indicators.